On 10/11/06, Robert McGwier <[EMAIL PROTECTED]> wrote: > Philip Covington wrote: > > On 10/11/06, Jim Lux <[EMAIL PROTECTED]> wrote: > > > >> At 06:44 AM 10/11/2006, Jim, W4ATK wrote: > >> > >>> Phil Covington wrote: > >>> > >>> "A lot of the perceived need for threading goes away when you get away > >>> from > >>> straight > >>> line procedural code." > >>> > >>> As an OLD machine language programmer, I am of the impression that > >>> "straight > >>> line" code increases the speed of the particular routine(s) avoiding those > >>> heavy far calls and the resultant stack operations. I think most > >>> optimizers > >>> that offer the choice of size and/or speed generate straight line objects > >>> as > >>> a means of increasing the speed..... Perhaps my rather ancient experience > >>> is > >>> mistaken.... > >>> > >> True in most cases. But, might be a case of optimizing for a > >> resource which we have an excess of, by spending a resource we are short > >> of. > >> > >> > > > > <snip> > > > > The interesting thing is that the JIT in .NET finds optimizations at > > run time that could not easily be built into an optimizing compiler. > > > > I was pleasantly surprised when I could compute a FFT/IFFT in code > > written in C# (based on Ooura's FFT code) as fast as using FFTW > > library (even after running wisdom). If you look through the SharpDSP > > source you'll see that I did not use FFTW. > > > > I suspect, but am not sure, that you ran these tests before we changed > the internals over to float and used the SIMD variants in FFTW. It is > very hard for me to imagine compiled code running faster than these hand > optimized assembly "codelets". That would be VERY interesting to > investigate if it is true. > > > > 73 de Phil N8VB > > > > _ > > 73's > Bob > N4HY
Actually, I have and its pretty much a wash for FFTs up to 65536 points (the largest I tested). I was trying to decide whether to use FFTW or the FFT in SharpDSP with the Mercury spectrum analyzer. I recompiled SharpDSP in .NET 2 and since the speed is pretty much the same I went with SharpDSP (which uses Ooura's FFT code). The big advantage is that I do not need to use another external library like FFTW. Also, in .NET 2 I removed all of the unsafe code with no performance hit. There really is no reason to use unsafe code in .NET 2. I am not surprised as much by the speed equivalence as I used to be when I realized that the JIT will use SIMD operations if it finds a performance gain. Now I realize that my C# coded DSP won't run on a TI DSP or dsPIC like DttSP can be modified to do, but it will run nicely on PCs in Windows and Linux (using mono). 73 de Phil N8VB _______________________________________________ FlexRadio mailing list FlexRadio@flex-radio.biz http://mail.flex-radio.biz/mailman/listinfo/flexradio_flex-radio.biz Archive Link: http://www.mail-archive.com/flexradio%40flex-radio.biz/ FlexRadio Homepage: http://www.flex-radio.com
